Xeons from Intel everywhere! - In Google datacenters with IPUs - With SambaNova RDUs for heterogeneous inference x86 seems stronger than ever, and one size does not fit all as always maintained. Meanwhile, Vikram Sekar is cross that Anthropic took away

MoE models are bandwidth-hungry. MatX claims to have more interconnect than any announced product, with a topology designed specifically for sparse MoE routing.

Guys, we're rich. Semi Doped is a hit podcast. Don't care what ya'll say. 😂

.Credo optics game is ON with Dust Photonics in the bag. Arista Networks XPO is a big deal that really deserves more attention... why? it is the last stand of the old world of optics vs the new fashioned CPO. Oh, and NUVACORE deserves some discussion, obviously!

"The best CPU for agentic AI is the one you can get your hands on." 🤣 -- Vikram Sekar

Meta's core business is ads. Ads are AI workloads. But not LLM workloads. @austinlyons chatted with Meta VP Matt Steiner to understand Meta's heterogeneous compute stack. Surprises: - Recommender training needs a different compute-to-memory ratio than LLMs. Hence MTIA. -

Seriously, listen to Matt Steiner explain this stuff. You wont find this info easily.

We just love when Google decides to put out such goodness just before our podcast recording day. We'll get you minty fresh takes only on Semi Doped. Stay tuned. blog.google/innovation-and…

CapEx week! Hyperscalers report Wednesday. semidoped.com/p/capex-week

Thanks to all our podcast listeners!

SiC in photonics? The ramp is fast.